From data to crypto to AI: three waves and what each one taught investors

Why look back at all

Every technology wave arrives with a story about where the value will go. The story is usually half right, and the wrong half is where investors lose money. I have evaluated more than 150 early stage startups as a venture operations analyst, and I build AI products myself. Both jobs keep pushing me toward the same question: when a cost curve bends, who ends up with the margin?

I want to answer that for three waves. Big data and cloud ran from roughly 2010 to 2017. Crypto and web3 ran from 2017 to 2022. Generative AI started in 2022 and is still running. The dates are loose. The pattern underneath them is not, and I think it tells us a lot about where AI value will settle.

Wave one: data and cloud

The cloud wave changed how computing was bought. Companies stopped buying servers and started renting compute and storage by the hour. Amazon did not even break out AWS revenue until April 2015, when it reported $1.57 billion for the first quarter, up 49% from a year earlier. Jeff Bezos called it a $5 billion business that was still accelerating. In 2024, AWS sales were $107.6 billion, with $39.8 billion of operating income.

That is the first place value went: to the incumbent with the balance sheet to build data centers and the patience to sell compute as a utility. As I read that era, many startups made a different bet. They packaged open source tools like Hadoop for enterprises that wanted to run their own clusters. The cloud absorbed a large part of that job. Cloudera, the company I most associate with that bet, agreed in June 2021 to be taken private by Clayton, Dubilier & Rice and KKR for about $5.3 billion, or $16 a share.

Now compare Snowflake. It sold a cloud data warehouse that runs on AWS, Google Cloud and Microsoft Azure, the same platforms that could have competed with it head on. Its IPO priced at $120 a share in September 2020, the largest software IPO ever at that point, and it ended its first day of trading worth about $70.4 billion. Snowflake did not own the cost curve. It rode the cost curve and owned the customer workload.

Investors got the direction right. The move to the cloud was real, and early bets on cloud native software paid. What they got wrong was treating data as a market when it was a layer of the stack, and funding companies whose main job was to package infrastructure that the hyperscalers would soon sell themselves. The lesson I carry forward: when a cost curve falls, value goes to whoever owns the workload and the customer, not to whoever repackages the commodity.

Wave two: crypto and web3

Crypto lowered the cost of something different: issuing and moving a digital asset without a trusted middleman. It also changed how startups raised money, because a token could be sold before a product existed.

The most influential thesis of the period was the Fat Protocols essay that Joel Monegro published at Union Square Ventures in August 2016. He argued that the internet had thin protocols and fat applications, and that blockchains would flip this, so value would concentrate at the shared protocol layer. For tokens he was partly right, since some base layer assets did hold very large market values. As businesses with cash flow, though, the winners looked a lot like the old financial system.

Coinbase, a regulated exchange, closed its first day on Nasdaq in April 2021 at a fully diluted valuation of about $85.7 billion, the largest direct listing up to then. Tether, a stablecoin issuer, reported $13 billion of net profit for 2024, based on an attestation by BDO. A company that holds more than $113 billion of Treasuries against dollars people want to use on a blockchain is running a bank like business, not a protocol.

Where investors went wrong was mostly timing and trust. Andreessen Horowitz announced a $4.5 billion crypto fund on May 25, 2022. On November 11, 2022, FTX filed for Chapter 11, its founder stepped down, and John J. Ray III took over to run the bankruptcy. My read is that many investors diligenced growth and token prices far more carefully than they diligenced controls. The lesson I carry forward: price is not usage, and counterparty and governance risk is part of the product, not a footnote.

Wave three: generative AI

Generative AI lowered the cost of producing language, code and reasoning. ChatGPT launched in November 2022 and reached an estimated 100 million monthly active users by the end of January 2023, according to a UBS note based on Similarweb data. TikTok took about nine months to reach the same number.

The curve that matters more for investors is price. In November 2024, Guido Appenzeller at a16z estimated that for a model of equal capability, inference cost was falling about 10x a year. GPT-3 level performance cost $60 per million tokens in November 2021 and about $0.06 three years later on a small open model. That is a 1,000x drop in three years.

The first large profits went to the chip maker. Nvidia reported fiscal 2026 revenue of $215.9 billion, with $193.7 billion from its data center business, up 68% from the year before. In June 2024, David Cahn at Sequoia called the gap between AI infrastructure spending and end user revenue AI's $600B question, and argued that GPU computing was turning into a commodity metered by the hour, with much less pricing power than physical infrastructure monopolies have.

The application layer is now catching up. Menlo Ventures estimated enterprise spending on generative AI at $37 billion in 2025, up from $11.5 billion in 2024. Applications took $19 billion of that, slightly more than the $18 billion spent on infrastructure, and startups earned nearly $2 for every $1 earned by incumbents at the application layer.

The pattern across all three

Put the three waves side by side and the same sequence shows up. The first dollars go to whoever sells the scarce input: data centers in the cloud wave, exchange access and block space in crypto, GPUs in AI. Then the input gets cheaper and more standard, and the margin moves toward whoever owns a repeatable customer workload on top of it.

Investors also keep overpaying for the middle. Hadoop distributions, many token projects and, I suspect, a large share of thin AI wrappers all sat between a commodity below and a customer above without owning either one. When the commodity got cheaper, the middle had nothing left to sell.

Trust compounds across every wave. I think AWS won because companies trusted it with production systems. Coinbase and the large stablecoin issuers, in my reading, won by sitting inside regulation rather than around it. In AI, my bet is that companies which carry real responsibility for outcomes in medicine, law and finance will be much harder to displace than companies that only generate text.

So the one lesson I carry into AI is simple. Falling model prices are a gift to applications and a threat to any company whose product is the model's output with a nicer interface. Underwrite the workload, not the model.

Where AI value accrues over the next five years

Here is my view, written so it can be checked against what happens by the end of 2031. First, the price of a fixed level of model capability keeps falling by at least 10x every two years, and model access starts to look like a utility with a few large suppliers. Second, application revenue grows faster than model API revenue and ends the period clearly larger, the way cloud software outgrew raw compute in the first wave. Third, inside applications, startups keep the majority of spend in workflow heavy verticals like healthcare, legal and financial services, while the model labs and the big software suites take most of the general assistant spend. Fourth, chip and data center suppliers stay very profitable, but their growth slows sharply as efficiency gains pile up.

If by 2029 the model labs' own products are taking most enterprise application spend, or token prices for frontier capability stop falling, I am wrong. In that world, the right portfolio is concentrated in the labs and the compute supply chain, and application investors are funding features for someone else's roadmap.